utils/prompts.py

"""
All LLM prompts used across CodeDebugger.
Centralised here for easy tuning and reproducibility.
"""

PROMPTS = {
    # ── Fixer agent ──────────────────────────────────────────────
    "system_fixer": (
        "You are an expert Python debugger. You will receive buggy Python code "
        "and an error message. Your job is to produce a FIXED version of the code.\n\n"
        "Rules:\n"
        "1. Return ONLY the corrected Python code inside a ```python ... ``` block.\n"
        "2. Do NOT add explanations outside the code block.\n"
        "3. Do NOT change function/class names or signatures.\n"
        "4. Do NOT add extra imports unless absolutely necessary.\n"
        "5. Fix the minimal number of lines required.\n"
        "6. Preserve existing comments and docstrings."
    ),

    "user_fixer": (
        "## Buggy Code\n"
        "```python\n{buggy_code}\n```\n\n"
        "## Error Message\n{error_message}\n\n"
        "## Previous Attempts (if any)\n{history}\n\n"
        "Fix the code. Return ONLY the corrected Python code in a ```python``` block."
    ),

    "user_fixer_with_feedback": (
        "## Buggy Code\n"
        "```python\n{buggy_code}\n```\n\n"
        "## Error Message\n{error_message}\n\n"
        "## Your Previous Fix\n```python\n{previous_fix}\n```\n\n"
        "## Test Results After Your Fix\n{test_results}\n\n"
        "## Feedback\n{feedback}\n\n"
        "Try again. Return ONLY the corrected Python code in a ```python``` block."
    ),

    # ── Critic / anti-hacking ────────────────────────────────────
    "system_critic": (
        "You are a code review critic. You verify that a proposed fix is "
        "legitimate and does not game the reward system.\n"
        "Check for:\n"
        "1. Hard-coded return values that match test expectations.\n"
        "2. Removal or neutering of test logic.\n"
        "3. Use of eval/exec to dynamically read expected outputs.\n"
        "4. Code that detects whether it is being tested.\n"
        "Return a JSON object: {\"is_suspicious\": bool, \"reason\": str}"
    ),

    "user_critic": (
        "## Original Buggy Code\n```python\n{buggy_code}\n```\n\n"
        "## Proposed Fix\n```python\n{fixed_code}\n```\n\n"
        "## Test Cases\n{test_cases}\n\n"
        "Analyze whether the fix is legitimate or is gaming the tests."
    ),

    # ── Reward explanation (optional, for UI) ────────────────────
    "reward_explanation": (
        "Given these reward components:\n{components}\n\n"
        "Provide a brief human-readable explanation of the agent's performance "
        "on this debugging step."
    ),
}

def get_fixer_prompt(buggy_code, error_type, description, test_cases, test_results=None, previous_explanation=None, iteration=1):
    prompt = (
        f"You are an expert Python debugger. Fix the following buggy code.\n"
        f"Bug description: {description}\n"
        f"Error type: {error_type}\n\n"
        f"Buggy Code:\n```python\n{buggy_code}\n```\n\n"
        f"Test Cases: {test_cases}\n"
    )
    if test_results:
        prompt += f"Previous Test Results: {test_results}\n"
    if previous_explanation:
        prompt += f"Previous Explanation: {previous_explanation}\n"
        
    prompt += (
        f"Iteration: {iteration}\n\n"
        f"Respond ONLY with a valid JSON object matching this schema:\n"
        f"{{\n"
        f'  "fixed_code": "the complete corrected python code",\n'
        f'  "explanation": "brief explanation of the fix"\n'
        f"}}\n"
    )
    return prompt

def get_simplified_prompt(buggy_code, error_type):
    return (
        f"Fix this {error_type} in the python code. Return ONLY the code in a markdown block, nothing else.\n"
        f"```python\n{buggy_code}\n```"
    )